By Peter Wayner, Originally posted on VentureBeat
When Hakim Weatherspoon looks at a cloud, he sees inefficiency everywhere. Sometimes the instances are overprovisioned and the extra RAM and CPU power costs money even when it’s not being used. Sometimes the code is running on a machine that’s too expensive. The cloud companies make it easy to dial up the power when there’s a burst of demand, but it’s not as easy to reduce it.
So he’s building tools to squeeze these inefficiencies out of the cloud. He’s the CEO and cofounder of Exostellar, an Ithaca, NY-based startup that’s building a layer of software that makes it simpler for running code to migrate to the cheapest hardware. It juggles containerized code by continually optimizing the performance, growing and shrinking the allocated hardware in response to demand. It’s designed to work across all clouds.
“The people who are in the most pain are the ones that are spending a ton on compute,” said Weatherspoon. “People who are spending a lot on simulations or other forms of high performance computing because they don’t have a good way of reducing their computational load.”
Lowering costs of computation
This week the company released the results from a study it finished with the Idaho National Laboratory. The project deployed Exostellar’s xSpot tool to orchestrate some of the lab’s computational experiments, which sometimes run for hours at a time. In the end, the bill for the computation dropped by more than 70%. The test studied the performance of MASTODON, a “high-performance computing simulation for structural dynamics, seismic analysis and risk assessment” that was packaged into a Docker container and run on AWS.
The savings came from running the project using machines rented from Amazon’s spot marketplace. All of the clouds run constant auctions for their spare computer cycles, and, in times of low demand, the savings can be 90% or greater.
“The spot market is cheap but not reliable,” said Weatherspoon. “We’re able to take the application as is and migrate it without stopping it into the spot market and then out [to standard instances.] It’s kind of a game of whack-a-mole without getting whacked.”
Exostellar has created a low-level virtualization technology that simplifies the migrations. xSpot watches factors like the current spot price for computing to predict when the cloud may shutdown an instance. When it seems like a shutdown of a spot machine is imminent, it moves the process and all of its state to a more expensive instance until the market clears.
“We have proprietary algorithms that take input of price and capacity and rebalance signals and a whole bunch of other things and we’ll trigger migration well ahead of time,” said Weatherspoon. “We’ve been so successful at it that we’ve been running in the spot for months actually for half a year. Without being terminated.”
The company also uses the same low-level techniques to repack running code into hardware in the most efficient way. xStack will consolidate running machines so that the minimum number of instances are used. xScale will dynamically adjust the size and power of the instances to work with current demand.
In some cases, the Exostellar layer is able to simplify running software by removing extra layers that may not be necessary for long-running code.
Giving you control in the cloud
“In this Idaho National Lab case, we are faster than their native implementation,” said Weatherspoon. “The virtual machine hypervisors provide isolation. We then have the operating system run at the same privilege level as the application. What that effect has is that now any system calls – any call from the application to the operating system – become a hundred times cheaper. “
The technology is designed to work alongside other popular container management options like Kubernetes, giving devops teams more options and control over cloud costs.
“When you go to the cloud, you don’t have [virtualization layers like hypervisors] anymore because Amazon or Google or Microsoft owns that physical machine and they own that software layer that virtualizes that machine you as a user don’t have access to,” explained Weatherspoon. “But our product does. It gives that access back to you.”
Exostellar maintains offices in Ithaca, New York and Santa Clara, California. Almost one year ago, Exostellar announced a $5m seed funding round led by Walden International and Nepenthe Capital LLC.
